Wegrw5: Agent Orchestration Workspace Based on GitHub Copilot

Wegrw5 is a research project focused on agent workflows within the GitHub Copilot ecosystem. It achieves intelligent task routing and autonomous execution via a multi-agent orchestration architecture, exploring new collaborative models for AI-assisted development. The project adheres to the 'deep dive into a single tool' philosophy—instead of pursuing cross-tool comparisons, it aims to maximize value within the capability boundaries of GitHub Copilot.

As GitHub Copilot evolves from a single code completion tool to a development platform supporting custom agents, how to effectively orchestrate multiple specialized AI agents has become a key issue for improving development efficiency. The Wegrw5 project was born in this context, building an intelligent workspace based on the GitHub Copilot ecosystem, focusing on multi-agent collaboration and task automation. Its uniqueness lies in the 'deep dive into a single tool' philosophy—it does not conduct cross-tool comparisons (e.g., with Aider, Claude Code, etc.) but focuses on unlocking value within Copilot's capabilities.

Wegrw5 adopts a 'command-execution' architecture, with @ben as the central orchestrator to coordinate specialized sub-agents for complex tasks. It draws on the enterprise project manager-expert team model to achieve efficient distribution and parallel processing. The tech stack is as follows:

Component	Technology	Description
Development Environment	VS Code	The only supported IDE (not Cursor or other tools)
Agent Runtime	GitHub Copilot CLI	Copilot binary for executing custom agents
Large Language Model	GitHub Copilot	Multi-model support (Claude Haiku 4.5, Claude Sonnet 4.6)
Memory System	Hindsight MCP	Semantic tags, observation scope, mental models
Agent Definition	VS Code Custom Agents	.agent.md format (YAML frontmatter metadata + Markdown instructions)

Wegrw5 defines 9 specialized agents, divided into two categories: entry agent (only @ben) and sub-agents:

• @ben (Orchestrator)：The only entry point of the system, responsible for analyzing user requests, decomposing tasks, and delegating to appropriate expert agents, following the five-step process of 'analysis-decomposition-delegation-coordination-reporting'.
• @doc (Documentation Expert)：Focuses on writing and maintaining technical documents (README, API docs, etc.). When a task involves documentation, @ben routes it here.
• @agentic-workflow-researcher (Workflow Research Expert)：Investigates agent workflow patterns, VS Code extensibility, etc., and provides cited professional analysis to support architectural decisions.
• @ar-director (HR Director)：When existing agents lack sufficient capabilities, designs and recruits new agents to expand the team.
• @ar-upskiller (Skill Enhancement Expert)：Follows the latest Copilot best practices and regularly updates agent definitions.
• @git-ops (Git Operations Expert)：Manages Git operations and enforces the Conventional Commits specification.

Wegrw5 introduces the 'hindsight' organizational memory system, building distributed intelligence through a four-stage progressive deployment:

1. Basic Memory: Establish a semantic tag system to organize discoveries, configure memory banks and handling rules, and implement basic retention/recall operations.
2. Knowledge Synthesis: Capture architectural patterns and best practices as mental models, and establish an automatic document update workflow based on research findings.
3. Intelligence Enhancement: Define observation scopes to filter discovery patterns, establish instructions for enforcing standards and norms, and inject hindsight capabilities into all 9 agents.
4. Emergent Intelligence: Form 24 mental models representing workspace knowledge, support reflection and synthesis of complex patterns, verify the effectiveness of production workflows, and continuously optimize them.

Example of a typical task processing flow:

1. User Initiates Request: Describe the requirement in Copilot (e.g., 'Add comprehensive API documentation for the payment service').
2. @ben Analysis and Routing: Identify the task as documentation writing and delegate it to @doc.
3. Delegation Execution: @ben passes the complete context to @doc, who independently researches the codebase and writes the documentation.
4. Result Reporting: @doc reports the list of modified files and key changes.
5. Commit Management: @ben can call @git-ops to handle commits and pushes, ensuring compliance with the Conventional Commits specification.

Independent tasks can be executed in parallel, while dependent tasks are orchestrated sequentially to maximize efficiency.

Wegrw5 follows the following design principles:

• Focus on a Single Ecosystem: All tools/models are limited to Copilot's capabilities to avoid cross-tool complexity.
• Intelligent Task Routing: The orchestrator analyzes request characteristics and accurately matches expert agents.
• Autonomous Execution: Expert agents work independently after receiving context, without the need for step-by-step human guidance.
• Parallel Workflows: Independent tasks are executed simultaneously to reduce overall time.
• Scalable Capabilities: Dynamically recruit new agents via @ar-director and continuously optimize existing capabilities via @ar-upskiller.

Practical Value: It verifies hypotheses such as specialized division of labor improving AI performance on complex tasks, explicit orchestration being superior to implicit stacking, the importance of organizational memory for long-term projects, and deep optimization of a single tool being comparable to multi-tool combinations. It provides reusable models for enterprise AI development teams in role design, task decomposition, and memory system construction.

Wegrw5 represents an important attempt in the evolution of AI-assisted programming toward 'teamization', demonstrating how a virtual team composed of specialized AI agents can collaboratively handle end-to-end tasks from research to operation and maintenance. As large language model capabilities improve and agent orchestration technology matures, similar architectures are expected to become a standard model for software development, redefining the boundaries of human-machine collaboration.